Exploring Agentic Visual Analytics: A Co-Evolutionary Framework of Roles and Workflows
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.
Tom: We're looking at the paper: "Exploring Agentic Visual Analytics".
Title and authors: Tom: We're starting with the title itself: "Exploring Agentic Visual Analytics: A Co-Evolutionary Framework of Roles and Workflows." It immediately tells you this isn't just about making a single AI tool better; it’s about a whole new way of structuring how we think about these systems.
Jane: I wonder what that co-evolutionary part means for us. It suggests that the agent roles aren't fixed; they change as the autonomy level goes up, and the human role changes too in response.
Lu: Right. They show this progression in Figure two which illustrates how agent roles evolve from Level one assistance to Level four strategic orchestration, and how human roles shift from direct command to high-level supervision > This suggests a dynamic relationship between the AI's power and the user's responsibility <ref:2604.15813#pg2,shift from direct command to high-level supervision>.
Meng: So if we build something at this level four autonomy, what does that mean for our operational stability? Are we talking about systems that can handle unpredictable tasks or is it still too brittle?
Lalam: The paper points out that as the system transitions from a reactive tool to a proactive one, there are new computational challenges. Specifically, they mention managing persistent analytical context and grounding multi-modal visual feedback > These are things you have to engineer carefully if you want it to work reliably in production.
The paper's summary: Tom: So the main takeaway is that agentic systems fundamentally reshape the traditional visual analytics pipeline. It’s not just a small tweak; it’s a complete overhaul of how data gets processed, mapped visually, transformed into views, and finally presented to us.
Jane: I'm hearing that the shift in data processing is moving from passive execution to proactive exploration and intelligent decision making across four dimensions. That sounds much more interesting than just running a pre-set script > What kind of exploration are we talking about there?
Lu: They argue that this means agents aren't just following rules anymore; they’re doing dynamic semantic alignment and advanced schema reasoning, moving away from needing fixed, predefined schemas for everything > This is key because real data is rarely perfectly structured.
Meng: That sounds computationally expensive. If the agent has to reason about the schema dynamically every time it processes data, how does that impact the cost per query when we scale up our users?
Lalam: They also cover visual mapping, where they move from simple, rule-bound chart generation to feature rich encodings driven by intent > This means the agent can handle complex structures and fine grained control over graphical properties based on what it’s trying to find.
The paper's improvements: Tom: Now let's talk about the suggested improvements they outline. They focus on making these systems more robust by addressing some of those inherent challenges, like how the agent actually performs its tasks when it gets complex.
Jane: I noticed a few key areas they suggest tackling. One is the CREATOR role, and they want to implement hybrid computation for that role > That sounds like a way to keep things reliable while still using the LLM’s high-level thinking power.
Lu: They specifically mention LEVA, which uses rule based methods for accurate tabular computation while using the Planner strictly for orchestration, avoiding those long delays you see in purely autonomous workflows like MisVisFix > It's about delegating deterministic tasks to explicit code so the LLM can focus on planning.
Meng: Hybrid computation sounds good for reliability, but what about transparency? If the Creator is just running code we don't see, how do we debug it when something goes wrong at scale?
Lalam: They suggest actionable transparency where systems provide explanatory feedback, like ReAct style thought logs, and allow humans to intervene to switch execution from parallel to serial > That lets you see the agent's reasoning path and even change where it’s going if needed.
Conclusion: Tom: So wrapping up, the paper argues that rising autonomy requires us to shift from being operators who execute commands to strategic supervisors who manage those high-level goals. It’s a partnership model now.
Jane: I think the big picture is that this framework gives us a concrete way to design these systems by defining those four roles and showing exactly where we need transparency and dynamic control > It moves the conversation from "can it do this?" to "how do we manage its autonomy safely?"
Lu: The paper emphasizes that for true high autonomy, we need profile aware agents that adapt to individual users over time through longitudinal learning frameworks > That’s where the next big research direction seems to be.
Meng: From an engineering standpoint, I'm still focused on the cost and deployment challenges. We need those mechanisms for dynamic autonomy allocation mentioned in the design guidelines so we can balance computational load against user expertise > If we don't do that, it just becomes too slow or too expensive for real traffic.
Lalam: The paper notes a limitation: they excluded early rule based systems and purely human centered design studies without novel technical systems > That means their framework might not apply perfectly to simpler, existing tools we use today.
Tom: So the big takeaway from "Exploring Agentic Visual Analytics: A Co-Evolutionary Framework of Roles and Workflows" is that while the potential for these agents is huge, we have to be careful about how much control we give up as the system gets smarter > It’s about balancing that trust.
Tianqi Luo, Leixian Shen, Yuyu Luo
cs.DB
Submitted: 2026-04-17
Updated: 2026-11-18
Code: https://github.com/marimo-team/marimo
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Importance score: 86/100
The gist: Agentic visual analytics (VA) represents an emerging class of systems in which large language model (LLM)-driven agents autonomously plan, execute, evaluate…
Key concepts
- PLANNER
- The PLANNER agent is responsible for high-level strategy and orchestration within the visual analytics pipeline. It translates a user's natural language goal into a sequence of necessary steps, managing the overall flow of the analysis and deciding which other agents should act next to achieve the final objective.
- CREATOR
- The CREATOR agent handles the generation of visual elements based on high-level intent. It moves beyond simple chart creation by performing feature-rich encodings, allowing it to handle complex data structures and exercise fine-grained control over graphical properties derived from deep analytical goals.
- Actionable Transparency
- This concept involves providing users with clear feedback and the ability to intervene in an agent's process. It means exposing the agent's internal reasoning—like thought logs—so humans can understand its decisions and switch execution paths, ensuring a collaborative rather than purely automated interaction.
- Metadata-First Abstraction
- This technique involves processing data by first summarizing it into structural metadata rather than using raw datasets. This allows agents to perform reasoning and generate execution scripts efficiently while maintaining privacy by keeping sensitive raw values local, supporting secure orchestration.
Terminology
Summary
Agentic visual analytics (VA) represents an emerging class of systems in which large language model (LLM)-driven agents autonomously plan, execute, evaluate, and iterate across the full visual analytics pipeline By shifting users from low-level tool operations to high-level analytical goals expressed through natural language, these systems are fundamentally transforming how humans interact with data
The Core Framework and Roles
The paper introduces a co-evolutionary framework that jointly analyzes the progression of agent autonomy alongside the necessary shift in human roles from manual operators to strategic supervisors Within this framework, four key agentic roles are defined: PLANNER, CREATOR, REVIEWER, and CONTEXTMANAGER This taxonomy maps these roles onto established VA pipeline stages and analyzes recurring trade-offs along three foundational axes: autonomy levels, agentic roles, and the VA workflow
Reshaping the Visual Analytics Pipeline
The introduction of agentic paradigm is not just an improvement in efficiency but a profound reshaping of the traditional visual analytics pipeline This examination covers how agentic visual analytics (AVA) disrupts and elevates the four classic stages—data processing, visual mapping, view transformation, and presentation
** Data Processing: From Static Schema to Autonomous Exploration:**
The stage shifts from passive execution to proactive exploration and intelligent decision-making across four key dimensions Agents perform dynamic semantic alignment and advanced schema reasoning, moving beyond reliance on predefined schemas
** Visual Mapping: From Simple Structures to FeatureRich Encodings:**
Agentic visual analytics elevates the complexity and flexibility of visual mapping, shifting from simple, rule-bound chart generation to feature-rich, intent-driven encodings This includes handling high-dimensional structures and exercising fine-grained control over graphical properties based on deep analysis goals
** View Transformation: Complex and High-Dimensional Navigation:**
Agentic systems introduce highly complex interaction paradigms and dynamic reconstruction capabilities, moving beyond manual brushing to intent-driven orchestration This includes intent-driven dynamic layout and reconstruction of entire dashboard layouts on the fly
Co-Evolution of Human-AI Interaction
As agent autonomy increases, human-computer interaction in visual analytics evolves from traditional command-based operations to a more collaborative, adaptive, and intelligent partnership This evolution is categorized into three primary dimensions: the evolving role of the human, the temporal dynamics of collaboration, and the modalities of control
** The Shifting Human Role:**
The user’s responsibility shifts from executing low-level tasks to managing high-level goals as agents take on PLANNER and CREATOR roles This shift requires enabling agents to handle sophisticated analytical tasks while ensuring users can follow the agent’s reasoning without being overwhelmed by conceptual burden
** Temporal Dynamics:**
Agentic systems transition to an iterative, persistent, and increasingly proactive temporal model This is enabled by maintaining temporal creation histories that allow users to conceptualize the analytical process and track session evolution
** Modalities of Control:**
Agentic visual analytics bridges the communication gap by integrating rich, multi-modal, and architectural interactions This includes combining natural language dialogue with direct spatial selections to ground the agent’s reasoning in the user’s visual focus
Design Guidelines for Trustworthy Adoption
The paper distills six actionable design guidelines derived from recurring design patterns and technical bottlenecks across the three taxonomic axes These guidelines address the fundamental trade-offs between agent autonomy and human-in-the-loop control
** Autonomy Levels:**
Design requires balancing computational costs, user expertise, and cognitive load by implementing dynamic autonomy allocation mechanisms that adapt to both the task’s inherent risk and the user’s proficiency level Designers must also employ non-disruptive proactivity using ephemeral UI elements to maintain cognitive flow
** Agentic Roles:**
Each role must be designed to mitigate the inherent limitations of large language models This involves hybrid computation for the CREATOR, offloading deterministic calculations to rule-based skills while reserving LLMs for high-level orchestration Actionable transparency is required for the PLANNER and REVIEWER to expose internal planning logic and allow users to modify execution paths
** VA Workflow:**
Systems must adapt how they handle data processing, visual mapping, and multi-modal interactions to ensure efficiency and security This includes operating on metadata summaries rather than raw datasets for privacy-preserving orchestration Agents must receive dual context—rendered visual outputs and underlying generation code—to ensure high-precision reasoning during view transformations
Limitations and Future Research Directions
Several limitations exist regarding the scope and methodology of the survey Scope exclusions include early rule-based systems and purely human-centered design studies without novel technical systems Methodological constraints arise from the rapid growth of multimodal AI, which may quickly introduce new paradigms Future research should focus on developing profile-aware agents that adapt to individual users over time through longitudinal learning frameworks Additionally, exploring Reinforcement Learning (RL) offers a pathway to true Level 4 autonomy by optimizing the PLANNER’s ability to dynamically restructure workflows The direction of GUI-Driven Visual Analytics involves positioning the agent as an autonomous digital assistant operating directly within existing GUIs via accessibility APIs Finally, research should explore generating Machine-Readable Visual Representations to facilitate the agent’s internal reasoning and self-correction
Conclusion
Our review of 55 systems reveals that LLMs enable decomposing complex visual analytics tasks into four specialized roles (PLANNER, CREATOR, REVIEWER, and CONTEXTMANAGER) Rising AI autonomy requires shifting the user from operator to strategic supervisor Our design guidelines emphasize dynamic autonomy allocation, actionable transparency, multi-modal synergy, and privacy-preserving evaluation Key challenges remain in visual detail blindness, orchestration latency, and human-AI alignment Addressing these gaps will be critical to realizing trustworthy, democratized agentic visual analytics >
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improve planning by implementing Hybrid Computation for CREATOR:
LEVA [73] uses rule-based methods for accurate tabular computation while using the PLANNER strictly for orchestration, avoiding the multi-minute delays seen in purely autonomous workflows like MisVisFix [10].
This ensures execution reliability by delegating deterministic tasks to explicit code, preserving the LLM’s cognitive overhead for planning. -
Improve transparency by implementing Actionable Transparency:
Systems must provide explanatory feedback (e.g., ReAct-style thought logs [71]) and allow humans to intervene, such as LightVA [72] enabling users to switch execution from parallel to serial.
This empowers users to override or steer the execution path by exposing agent logic through interactive interfaces. -
Improve memory by implementing Persistent Contextual Memory for CONTEXTMANAGER:
Effective memory management, such as a global TODO list [7] or “contextual interaction linking” [4], prevents agents from losing track of the user’s spatial focus.
This ensures cross-agent consistency by acting as a robust, centralized module that synchronizes state and preserves dialogue history across all agents. -
Improve data processing by implementing Metadata-First Abstraction in Data Processing:
CoDA [7] advocates profiling data into structural summaries, allowing the PLANNER and REVIEWER to generate execution scripts without accessing raw values.
This enhances efficiency and supports privacy-preserving orchestration by keeping sensitive data local while enabling reasoning based on summaries. -
Improve visual mapping by implementing Hybrid Contextual Grounding in Visual Mapping:
ChartEdit [70] proves that systems using only visual perception struggle with data precision compared to those grounded with both chart and code.
Agents must receive this dual context to ensure high precision reasoning and preserve underlying data logic during view transformations. -
Improve interaction by implementing Multi-modal Synergy in Interaction:
Systems should adopt a synergistic approach where linguistic intent is combined with spatial gestures (e.g., lassoing) [4], while also interpreting implicit behaviors like hovering.
This blends natural language for abstract intent with direct GUI manipulation for precise spatial selection, keeping the agent’s mental model continuously synchronized with the user’s visual focus.
Abstract
Agentic visual analytics (VA) represents an emerging class of systems in which large language model (LLM)-driven agents autonomously plan, execute, evaluate, and iterate across the full visual analytics pipeline. By shifting users from low-level tool operations to high-level analytical goals expressed through natural language, these systems are fundamentally transforming how humans interact with data. However, the rapid proliferation of such systems in recent years has outpaced our understanding of their design landscape. Two intertwined problems remain open: how do autonomous agents reshape the traditional VA pipeline, and how must human involvement adapt as agent autonomy increases? To address these questions, this paper presents a comprehensive survey of 55 primary agentic VA systems and introduces a co-evolutionary framework. This framework is essential because it jointly analyzes the progression of agent autonomy alongside the necessary shift in human roles from manual operators to strategic supervisors. Within this framework, we define a role-workflow taxonomy that aligns four key agentic roles (PLANNER, CREATOR, REVIEWER, and CONTEXT MANAGER) and maps them onto established VA pipeline stages. Our analysis uncovers recurring trade-offs along three foundational axes: autonomy levels, agentic roles, and the VA workflow. We consolidate these findings into actionable design guidelines and outline future research directions for agentic visual analytics. A web-based interactive browser of our co-evolutionary framework, including the corpus and design guidelines, is available at agenticva.github.io/AgenticVA/.
Sources
- Large Language Model based Multi-Agents: A Survey of Progress and Challenges
- Conversational AI Threads for Visualizing Multidimensional Datasets
- LLM-Assisted Visual Analytics: Opportunities and Challenges
- DataNarrative: Automated Data-Driven Storytelling with Visualizations and Texts
- Prompt4Vis: Prompting Large Language Models with Example Mining and Schema Filtering for Tabular Data Visualization
- Facilitating Conversational Interaction in Natural Language Interfaces for Visualization
- Text2Vis: A Challenging and Diverse Benchmark for Generating Multimodal Visualizations from Text
- Generating Analytic Specifications for Data Visualization from Natural Language Queries using Large Language Models
- Natural Language Interfaces for Tabular Data Querying and Visualization: A Survey
Related papers
- Beyond Overlap: Hierarchical Graph Abstraction for Out-of-Core Vector Search
- EMA: Approximate Nearest Neighbor Search with General Attribute Filtering and Dynamic Updates
- Open Biomedical Knowledge Graphs at Scale: Construction, Federation, and AI Agent Access with Samyama Graph Database
- Samyama: A Unified Graph-Vector Database with In-Database Optimization, Agentic Enrichment, and Hardware Acceleration
- GenTUS: Table Union Search via Constrained Generation of Semantic Table Identifiers
- Common-Neighbor-Count-Based Representative Possible World Finding on Uncertain Graphs